6.s062: mobile and sensor computing
TRANSCRIPT
6.S062: Mobile and Sensor Computing
Lecture4:Device-FreeLocaliza3on
So Far: Device-based Localization
This Lecture: Using radio signals to track humans without any sensors on
their bodies
This Lecture: Using radio signals to track humans without any sensors on their bodies
Operates through occlusions
• Location • Vital Signs • Gestures
Example: WiTrack
Device in another room
Device
ApplicationsSmartHomes
EnergySaving Gaming&VirtualReality
Measuring Distances
Rx
Tx
Distance = Reflection time x speed of light
Measuring Reflection Time
Time
Txpulse Rxpulse
Option1:Transmitshortpulseandlistenforecho
ReflectionTime
Why?
Measuring Reflection Time
Time
Txpulse Rxpulse
Option1:Transmitshortpulseandlistenforecho
Capturingthepulseneedssub-nanosecondsampling
SignalSamples
ReflectionTime
Why was this not a problem for Cricket?
Capturingthepulseneedssub-nanosecondsamplingWhy?
Multi-GHz samplers are expensive, have high
noise, and create large I/O problem
Time
Freq
uenc
yTransmitted
t
FMCW: Measure time by measuring frequency
Time
Freq
uenc
yTransmitted
t t+ΔT
Received
FMCW: Measure time by measuring frequency
Reflection Time
How do we measure ΔF?
ΔF
ΔF slope=
Measuring ΔF
MixerTransmittedReceived
Signal whose frequency is ΔF
FFT
Power
ΔF
• Subtracting frequencies is easy (e.g., removing carrier in WiFi)
• Done using a mixer (low-power; cheap)
Measuring ΔF
MixerTransmittedReceived
Signal whose frequency is ΔF
FFT
Power
ΔF
ΔF ➔Reflection Time ➔ Distance
• Subtracting frequencies is easy (e.g., removing carrier in WiFi)
• Done using a mixer (low-power; cheap)
Challenge: Multipath➔ Many Reflections
Rx
Tx
Distance
Refle
ctio
n Po
wer Multi-paths mask person
Static objects don’t move ➔ Eliminate by subtracting consecutive measurements
Distance
Power
Distance
Power @timet+30ms
@ time t
-=
Multi-path
Multi-path
2meters
Power
Distance
Why 2 peaks when we only have one moving person?
Challenge: Dynamic Multipath
Rx
Tx
Distance
Power
DynamicMulti-path
MovingPerson
The direct reflection arrives before dynamic multipath!
Mapping Distance to Location
Person can be anywhere on an ellipse whose foci are (Tx,Rx)
By adding another antenna and intersecting the ellipses, we can localize the person
TxRx
d
Rx’
From Location to tracking
Challenge: Dynamic Multipath
Rx
Tx
Distance
Power
DynamicMulti-path
MovingPerson
The direct reflection arrives before dynamic multipath!Fails for multiple people in the environment, and we need a more comprehensive solution
How can we deal with multi-path reflections when there are multiple
persons in the environment?
Idea: Person is consistent across different vantage points while multi-path is different
from different vantage points
Combining across Multiple Vantage PointsExperiment: Two users walking
Setup Single Vantage Point
Mathematically: each round-trip distance can be mapped to an ellipse whose foci are the transmitter and the receiver
Experiment: Two users walkingSetup Two Vantage Points
Combining across Multiple Vantage Points
Experiment: Two users walkingSetup 16 Vantage Points
Localize the two users
Combining across Multiple Vantage Points
Multi-User LocalizationExperiment: Four persons walkingSetup All Vantage Points
four persons
-4 -3 -2 -1 0 1 2 3 4Distance (meters)
0 1 2 3 4 5 6 7 8
Dis
tanc
e (m
eter
s)
first person other people or noise?
Multi-User LocalizationExperiment: Four persons walkingSetup All Vantage Points
four persons
-4 -3 -2 -1 0 1 2 3 4Distance (meters)
0 1 2 3 4 5 6 7 8
Dis
tanc
e (m
eter
s)
first person other people or noise?
Near-Far Problem: Nearby persons have more power than distance reflectors and can mask them
Successive Silhouette Cancellation: a new algorithm that localizes multiple
persons in the scene by addressing the near-far problem
Decode human location
Goal: Recover human reflections
Model human and reconstruct
reflection patterns
Subtract
Successive Silhouette Cancellation: a new algorithm that localizes multiple persons in
the scene by addressing the near-far problem
First localize the user with the strongest reflection
-4 -3 -2 -1 0 1 2 3 4Distance (meters)
0 1 2 3 4 5 6 7 8
Dis
tanc
e (m
eter
s)
Person 1 at (-0.5, 1.3)
Cancel the impact of the person’s whole body
After reconstructing and cancelling the first user’s reflections
-4 -3 -2 -1 0 1 2 3 4Distance (meters)
0 1 2 3 4 5 6 7 8
Dis
tanc
e (m
eter
s) Person 2 at (0.5, 2)
-4 -3 -2 -1 0 1 2 3 4Distance (meters)
0 1 2 3 4 5 6 7 8
Dis
tanc
e (m
eter
s)Iteratively localize the remaining users in
the scene
Person 3 at (0.8, 2.7)
-4 -3 -2 -1 0 1 2 3 4Distance (meters)
0 1 2 3 4 5 6 7 8
Dis
tanc
e (m
eter
s)Iteratively localize the remaining users in
the scene
Person 4 at (1, 4)
How can we localize static users?
Dealing with multi-path when there is one moving user
Rx
Tx
We eliminated direct table reflections by subtracting consecutive measurements
Needs User to Move
Dealing with multi-path when there is one moving user
Rx
Tx
STATIC
We eliminated direct table reflections by subtracting consecutive measurements
Needs User to Move
Exploit breathing motion for localize static users
• Breathing and walking happen at different time scales –A user that is pacing moves at 1m/s –When you breathe, chest moves by few mm/s
• Cannot use the same subtraction window to eliminate multi-path
User Walking at 1m/s
30ms subtraction window 3s subtraction window
Localize the person
Person appears in two locations
User Sitting Still (Breathing)
Cannot localizeLocalize the
person
-4 -3 -2 -1 0 1 2 3 4Distance (meters)
0
1
2
3
4
5
6
7
8
Dis
tanc
e (m
eter
s)
30ms subtraction window 3s subtraction window
User Sitting Still (Breathing)
CannotlocalizeLocalizetheperson
-4 -3 -2 -1 0 1 2 3 4Distance (meters)
0
1
2
3
4
5
6
7
8
Dis
tanc
e (m
eter
s)
30ms subtraction window 3s subtraction window
Use multi-resolution subtraction window to eliminate multi-path while being able to localize
both static and moving users
Localize the two users
Centimeter-scale localization without requiring the user to carry a wireless device
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2
y-ax
is (m
eter
s)
Localize the two users
People are points
Want a silhouette
Approach: Combine antenna arrays with FMCW to get 3D image • 2D Antenna array gives 2 angles • FMCW gives depth (1D)
2D array
1D
1D
49
Output of 3D RF Scan
Blobsofreflectionpower
Challenge: We only obtain blobs in space
Cannot Capture Reflection
Challenge: We only obtain blob in spaceAt frequencies that traverse walls, human body parts are specular (pure mirror)
RF Scanning
Setup
At every point in time, we get reflections from only a subset of body parts.
Solution Idea: Exploit Human Motion and Aggregate over Time
RF Scanning Setup
Solution Idea: Exploit Human Motion and Aggregate over Time
RF Scanning Setup
Previous Location
New Location
Combine the various snapshots
3m 2.5m 2m
Chest (Largest Convex Reflector)
Use it as a pivot: for motion compensation and segmentation
Human Walks toward Sensor
3m 2.5m 2m
Chest (Largest Convex Reflector)
Use it as a pivot: for motion compensation and segmentationCombine the various snapshots
Right arm
Left arm
Head
Lower TorsoFeet
Human Walks toward Sensor
Human Walks toward Sensor
SampleCapturedFiguresthroughWalls
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2
y-ax
is (m
eter
s)
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2
y-ax
is (m
eter
s)
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2y-
axis
(met
ers)
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2
y-ax
is (m
eter
s)
Sample Captured Figures through Walls
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2
y-ax
is (m
eter
s)
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2
y-ax
is (m
eter
s)
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2y-
axis
(met
ers)
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2
y-ax
is (m
eter
s)
SampleCapturedFiguresthroughWalls
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2
y-ax
is (m
eter
s)
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2y-
axis
(met
ers)
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2
y-ax
is (m
eter
s)
-0.2 0 0.2 0.4 0.6 0.8x-axis (meters)
0
0.5
1
1.5
2
y-ax
is (m
eter
s)
Sample Captured Figures through WallsThrough-wall classification accuracy of 90% among 13 users